7.4 Default Credibility Learning
7.5.3 Default credibility learning
We now turn to the evaluation of the last learning algorithm introduced in this chapter. This algorithm is designed to monitor the general level of credibility of witnesses or referees in an environment and to update the default witness/referee credibility value. Hence, the effectiveness of this algorithm is going to be tested in situations where the general level of lying in the environment changes. In this experiment, there are two main groups of consumer agents; both of them using the CR component with the credibility model extension as their trust model. However, one of them is equipped with the default credibility learning algorithm, calledCRL.
-20% 0% 20% 40% 60% 80% 100% 120% 1 21 41 61 81 101 121 141 161 181 Round Ly in g pe rc en ta ge
Figure 7.11: Learning TDRCr— CR lying percentage.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 21 41 61 81 101 121 141 161 181 Round
Default Referee Credibility
Figure 7.12: Learning TDRCr—TDRCr’s evolution.
The other group is calledCR. Since the default credibility is only used for referees that are met for the first time, new referees need to be continuously added into the testbed in order to evaluate this facet of the model. Therefore, we introduce a dummy group of consumers whose tasks are simply interacting with the providers and providing references for them. All of the consumers in this group are lying referees of eitherExtr1or Extr2type. Similar to Section 7.5.1, the level of lying in the testbed is controlled by using the CR lying percentage, which is set following the chart in Figure 7.11. To ensure the abundance of new referees in every round, 30% of the agents in the dummy group are replaced by new dummy consumer agents after each round. Since the dummy consumers do not use a trust model (they randomly select providers for interaction), we only monitor the performance of CRand CRL.
The experiment’s results are presented in Figures 7.12 and 7.13. The TDRCr value of groupCRL, plotted through time in Figure 7.12, closely corresponds to the CR
0 0.51 1.52 2.53 3.54 1 21 41 61 81 101 121 141 161 181 Interaction P erf orman ce 0 1 2 3 4 5 6 Ran k CR CRL R.CR R.CRL
Figure 7.13: Learning TDRCr performance.
lying percentage in Figure 7.11. Thus, TDRCr is raised to its near maximum value (i.e. 1.0) when the CR lying percentage is minimum and vice versa. It is apparent here that the default credibility learning algorithm can effectively track the level of lying/inaccuracy in the environment and update the default referee credibility accordingly. Figure 7.13 shows that, with the new learning algorithm, CRL can outperform CR in several instances while it has similar performance with CR in the other cases. The slight performance improvement is expected because TDRCr is only used for new referees and, thus, it cannot significantly affect the overall performance of FIRE.
7.6
Summary
This chapter has further extended FIRE towards a more flexible and adaptable trust model. We have explored several learning techniques and partly automated the process of choosing the right parameters in order to ensure that FIRE oper- ate effectively in a range of environments. Specifically, and most importantly, we devised an algorithm to monitor the performance of each of FIRE’s components (based on the accuracy of their trust values). This is a novel approach that al- lows any changes in an agent’s environment that affect the performance of one or more components (e.g. lack of ratings or changes in the quality of ratings from a particular source of information) to be indirectly detected and the weights for the corresponding components (i.e. the component coefficients) to be accordingly adjusted. In addition, appropriate algorithms are also devised for choosing the right values for the inaccuracy tolerance threshold and the default witness/ref- eree credibility. When taken together, all these new techniques enhance FIRE’s
robustness and resilience in facing unforeseen circumstances. Through empiri- cal evaluation, we have shown that the new algorithms are in fact effective and significantly improve FIRE’s adaptability.
This chapter has concluded the research work of designing a generic and adaptable trust and reputation model for applications in open MAS in the scope of this thesis. The next chapter will summarise the contributions this research has made and outline the directions for the future work.
Conclusions
T
hischapter summarises the findings of this thesis in enabling agents in an openmulti-agent system to assess the trustworthiness of their peers for selecting good interaction partners. In order to do so, a novel trust and reputation model — FIRE — was developed, which takes into account the main characteristics of an open MAS to ensure its robustness and applicability in such environments. More specifically, this thesis presents and evaluates a framework for evaluating trust- worthiness of agents based on multiple sources of information: direct experience, role-based relationships, witness reports, and certified references. By using this framework, agents in an open MAS are able to obtain the trust values of their peers in most circumstances. This is possible because FIRE does not rely solely on one source of information (its four components complement and back up one another) and particularly because it exploits the high availability of the novel Certified Reputation component.
In more detail, Section 8.1 reviews the contributions of this research to the state of the art. Section 8.2 then discusses the main ways in which this research can be carried forward in the future.
8.1
Research Contributions
This thesis has presented FIRE, a novel decentralised model for trust evaluation that is specifically designed for general applications in open MAS. Before going on to FIRE’s contributions to the state of the art, we recap the requirements for a trust model for applications in open MAS (discussed in Section 2.5):
R1a It should deal with the bootstrapping issue of newly joined agents.
R1b It should make use of role-based trust, interaction trust, and witness reputa- tion when the required information for these dimensions of trust is available.
R2a Each agent should be able to collect observations and calculate the reputation values by itself.
R2b The trust model should be scalable to a large number of agents that might be present in open MAS.
R2c It should reasonably maintain its normal effective operation in situations where there are various changes in its environment.
R3 It should be adaptable to different domains of applications that an open MAS may have.
R4a It should be robust against possible lying from agents and
R4b the correlated evidence problem.
In what follows, we are going to show how these requirements are met by FIRE and highlight its novelties. In an overview, the novel mechanisms developed in this research can be classified into the following areas:
• evaluating trust: A generic framework is built which allows a variety of sources of trust information to be integrated to provide a collective and precise trust measure. The model is able to predict closely the behaviour of an agent. In addition, Certified Reputation, a novel type of reputation, is introduced. (Chapter 3)
• dealing with inaccurate reports: A model of the reporter’s credibility is de- veloped, allowing FIRE to weight third-party reports according to their provider’s credibility and filter out inaccurate reporters. (Chapter 6)
• adapting to the environment: Learning techniques were implemented to adapt a number of FIRE’s parameters to the prevailing context, allowing it to operate robustly under unforseen circumstances in the environment. (Chapter 7)
The remainder of this section discusses the results of the work in this thesis focusing on the above aspects in turn (in Sections 8.1.1, 8.1.2, and 8.1.3, respectively).